The Data and Images of Natural Disaster News

Report Based on Artificial Intelligence Technology

Processing

hunxiao Wang

Computer Science

City University of Hong Kong

Beijing, China

cx894448636@163.com

Ruiqi Peng

Data Journalism

Communication University of China

Beijing, China

2878572825@qq.com

bstract—In recent years, the technology of writing data news

report by using artificial intelligence has been rapidly developed.

The news that is generated automatically has strong timeliness.

However, due to lack of humanistic care, it can't be widely used in

long reports. Based on the collection and analysis ability on data

of artificial intelligence, the data news can be completed by

program algorithms and machine learning. By taking use of

natural disaster data extracted by crawler, and achieving

recognition and classification of disaster weather warning images

by convoluted networks, this paper comes up with the model of

data and images in the natural disaster news report based on

artificial intelligence technology processing.

Keywords—artificial intelligence, natural disaster, data, images

I. I

RODUCTION

. Research Background

At 21:19 on August 8, 2017, a magnitude 7.0 earthquake

occurred in Jiuzhaigou County, Aba Prefecture, Sichuan

Province. Within 18 minutes after earthquake occurrence, the

first news “a magnitude 7.0 earthquake occurred in Jiuzhaigou

County, Aba Prefecture, Sichuan Province” was pushed by

WeChat official account “China Seismological Network”, of

which the contents contained quick report parameters, epicenter

topography, thermal population, historical earthquake, epicenter

weather and others. This report was generated by computer

taking 7 seconds to complete. Pan Huaiwen, the director of the

China Earthquake Network Center, pointed out that “the robot

enters into push platform, the spread speed would be fast and

accurate, which can cover the major people within several

seconds after the earthquake information report is completed”

[1].

he cases of news report written by artificial intelligence

have not been rare. The news generated automatically has strong

competitive force in timeliness and accuracy, becoming the

natural advantages of developing to write news report by

artificial intelligence. However, currently, the artificial

intelligence fails to play its role in long report due to lack of

humanistic care and emotional intention, which are mainly

applied for news alerts. At the same time, the acquisition and

analysis ability on data of artificial intelligence coincides with

the demand of data news. As a consequence, this paper would

like to discuss the feasibility of artificial intelligence in the

application to data news field, and bring up with the applicable

model of data and images in natural disaster news report

processed by artificial intelligence.

B. Journals Reviewed

According to a study by Guan et al. [2] at China University

of Geosciences in 2021, it is shown that artificial intelligence

has been focused on solving seismic data processing and

comprehensive interpretation in the field of fluid deposit

exploration. In addition, it has been widely used in automatic

rock lithology and mineral classification, geological hazard risk

assessment, intelligent interpretation of remote sensing images,

natural seismic signal analysis and monitoring and prediction.

And a variety of machine learning and intelligent computing

methods have been able to meet the computing speed and

prediction accuracy required by the large volume of seismic

wave data, such as the combination of remote sensing image

processing, neural network, genetic algorithm and swarm

intelligence optimization algorithms to improve the image edge

recognition accuracy.

In the article “AI and IoT in Atmospheric Sciences”, it is

shown that “three types of intelligent analysis can be achieved

using the ‘AI+IoT’ model”, namely real-time analysis, optimal

analysis and predictive analysis. “Meteorological researchers

are constantly trying to use this model for weather and climate

prediction, weather disaster warning, etc. The application of this

field is currently in its infancy, but in the future the integrated

use of AI and IoT has great potential and development space in

the field of meteorology.” Currently, this technology is mainly

applied in the fields of meteorological observation identification,

meteorological data processing, and weather and climate

analysis and forecasting [3].

According to Zhou's research [4], the current artificial

intelligence method can realize the three-dimensional analysis

and expression of pre-disaster warning, which provides a new

technical method for pre-disaster warning, and “disaster

management needs to capture real-time social media big data,

and the development and application of specific crisis

classification, entity classification and data aggregation

technology is urge, and there is also a need to present and

visualize social big data through maps”, and “artificial

intelligence has the characteristics of virtualization,

The

se authors contributed equally to this work

227

2023 15th International Conference on Computer Research and Development

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 08:10:07 UTC from IEEE Xplore. Restrictions apply.

ontextualization and science, which is one of the best

technologies for post-disaster recovery” [4].

From the above literature, we can see that AI has played a

great role in the research related to natural disasters, and the

current technological development can fully support its

academic research and application. It is undeniable that the data

collection methods in the above research are worthy of our

research in this subject.

However, these research topics still focus on climate

prediction, pre disaster warning, etc., and do not visualize the

data in news reports. Few studies have combined the two and

conducted research on this subject.

The model in the above research is not applicable to news

reporting. The two most obvious problems are:

• The images produced are too academic. The audience of

news reports is far more than that of researchers

engaged in natural disasters, so it is necessary to

describe natural disasters in the most intuitive and

concise words and images. High reading threshold will

directly reduce the communication efficiency.

• The production process is too slow. The pursuit of

timeliness in news reporting, which spends too much

time on highly professional analysis, is not conducive to

high-quality communication.

In conclusion, the author believes that in the field of natural

disaster reporting, in order to facilitate understanding, ensure

accuracy and improve output efficiency, AI needs to reduce

professional analysis and use visual means of data visualization.

“The information obtained by the audience will have a greater

impact on the audience because of the preconception. The lack

of information caused by delayed reporting or even concealment

will leave room for rumors to spread. In serious cases, it will

even make the media lose the ability to set the agenda, causing

the spread of panic, reducing the credibility of the government,

and causing secondary harm to the society” [5]. When major

events such as natural disasters occur, news reports should not

be silenced at the new media end, but should highlight the

mobile priority strategy, and launch faster, more and better news

reports and financial media works at the “two micro end” [6].

To solve the above problems, we choose to use the crawler

algorithm to obtain public natural disaster related data.The

natural advantage of open data is that it is less professional and

easy to understand. Crawler algorithms can provide efficient

data retrieval for data visualization, speed up the production of

reports, and thus provide highly timely reports for the public.

II. I

RODUCTION OF

. Multiplicity of Data Semantics

1) Data Storage

As it’s shown in figure 1, data exists as a symbol, and symbol

has its own semantics. Data can store ancient and past

information, preserving the instantaneous moments that have

been lost and helping us to track history. This is also the reason

for data can be compared for time. Meanwhile, the characteristic

of data can help innovation because innovation needs to focus

on previous experience. For example, the record of data museum

about historical dynasties can reflect this characteristic. In the

natural disaster report, the storage of data can help media and

the public to obtain past data so as to briefly summarize trends

to form a collection.

2) Fragmentation of Data

The data is divided into different segments by structure.

Each segment becomes the information carrier of different

topics or views that empowering data diversity. Due to the

fragmentation of data, it needs to guarantee the used data are

convenient for communicating, solving and dealing with when

using it. Meanwhile, when reporting natural disaster, this

characteristic can enhance the logic of report and describe this

disaster from multiple dimensions.

3) Abstractness of Data

Data is not the completed retention of the initial information.

The public can't see or feel some information through data,

instead they need to recover, process and use these scenarios

through retained data by thinking. Data refines the essential

characteristics from various matters of this event, and eliminates

the non-essential characteristics. Just because of this feature,

news report can extract concrete images from fact. As ancient

documents recording, it is impossible to record all figures, all

matters at all views, but to record their core and structure.

4) Granularity of Data

The granularity of data is similar to the pixels of pictures that

countless of granules composing the data and carrying

information. Meanwhile, the bigger of granularity is, the smaller

the information carried by the data, and the easier the data to be

understood. This characteristic of data is relevant with hierarchy

of natural disaster reports. For example, in the same natural

disaster event, some data only record the year of occurrence, but

some data accurate all events into seconds.

B. Journalists' Subjectivity: Two Paradigms

At present, there are two main paradigms of data news report.

Data driving news is that ones driven by data processing and

discussed from multiple technical perspectives. “Data driving

news is the tool to analyze and filter huge amounts of news data.

It digs out news through integrating data.” Data can drive media

field to produce contents, such as judging whether movies are

successful or not by scientific calculation. There are many risky

factors in the data driving news, such as the different quality of

data, the restriction of journalist’s subjectivity, and the

ignorance of relevance. The data are more emphasized in the

data driving news, while the news is restricted. Data supports the

narrative, but can't complete the narrative. Overemphasizing

data will lead the increment of the narrative results, and the role

of journalist editing will decrease.

The news report assisted by data has the feature of taking

data as foundation and assistance. News is still the main body,

of which the essence is “narrative drive” based on data analysis,

and the data is just the means and methods of narrative

assistance.” Telling a good story is still the ultimate and

fundamental demand of data news.” “Journalists must be the

managers of data.” The CAR emerged in the early time

(computer assisted reporting, investigation, reference and

gathering talking) has been gradually replaced by data assisted

news reporting along with the coming of ABC (AI, Big data,

228

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 08:10:07 UTC from IEEE Xplore. Restrictions apply.

ig. 1. Multiplicity of data semantics

rong convective weather data collection

obtain relevant data of

new strong convective

weather

Recognition and

acquisition of strong

convective information

he data of strong convective weather collected by this crawler

Store URL and

match it with

database

nitial URL

new URL

integrate and classify

images and data of

strong convective data

terminate crawler

ake judgement about that whether terminate

crawlerbased on the region,information

completionand other comprehensive conditions

ig. 2. The acquisition process of crawler’s intelligent disaster data

loud). To sum up, the main difference between two kinds of

data news report paradigms is the driving force. The data driving

news report is driven by data, while that of data assisted news

report is driven by narrative. The priority of “data” and “news

narrative” is different, which leading to more different points

mentioned as above. Currently, due to journalists are still

playing the main role in news reporting, the data assisted news

report has a more common paradigm in application.

III. A

PLICATION OF

TIFICIAL

TELLIGENCE IN

URAL

SASTER

PORT

ELD

or natural disaster related data, there are two main parts,

one is textual information data and the other is image data. One

of the main features of textual information data is that it is

relatively standardized relative to textual and professional data,

such as strong convective weather and the corresponding level,

dry weather and the level of drought degree. However, the

relative types are more complicated. Taking weather disasters as

an example, the typical ones are strong convection weather,

morning fog weather, typhoon weather, drought weather,

geological disaster weather, etc. The classification of each

disaster weather level has its own characteristics. On the other

hand, the image data is characterized by the fact that most of

them have the map templates of sea and land boundaries, and the

corresponding colors are used to distinguish different levels for

different kinds of weather hazards.

Based on the analysis of the data characteristics related to

natural disaster news, we also have a corresponding strategy in

the relevant processing algorithm, as follows.

1) Crawler algorithm: Taking weather disasters as an

example, most of the related warnings and real-time

data come from professional public portals on the

Internet. On the other hand, the corresponding

professional names of weather disasters are more fixed

and keyword retrieval is easier, so we can use crawler

algorithm to obtain real-time and wide range of first-

hand related data information.

Using crawler algorithm to obtain public natural disaster

data has the following advantages:

• Improve the efficiency of natural disaster data retrieval：

In the era of information explosion, data information is

complex and highly timely. This is especially true of

public data related to natural disasters. For example,

when a typhoon occurs in a certain area, yo

retrieve typhoon information on various portals,

professional websites, social platforms, etc., but if you

want to obtain accurate, open and authoritative target

disaster data at the first time, it will undoubtedly bring

huge workload. However, the use of the crawler

algorithm will greatly reduce the labor cost, and can

229

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 08:10:07 UTC from IEEE Xplore. Restrictions apply.

ore efficiently and accurately obtain the relevant data

of the target natural disasters.

• Crawlers can make data acquisition more accurate：

The data of natural disasters contain a lot of professional

data and technical content. Many of the data content is

often not the material required by the press release, so

other professionals are required to conduct manual

identification and screening. However, the quality of the

news data that are subject to manual identification and

screening is not uniform. After using the crawler, we

can accurately screen the natural disaster data we need

after formulating standard screening rules.

2) Convolutional neural network recognition and

classification. Since natural disaster data will have

many image data, such as typhoon path map and

distribution map of strong convective weather, we can

apply the convolutional neural network of typical image

recognition to train the recognition of disaster data

images, or take the strong convective weather in

weather disaster as an example, on the one hand, we can

identify the type of disaster by the title name in the

image, and on the other hand, we can identify different

levels of disaster by the training color.

Specifically, the convolution algorithm has the

following advantages：

• Recognition of image data that can deal with natural

disasters: We all know that convolutional neural

networks are widely used in image recognition, speech

recognition and other fields. For natural disasters,

images are one of the most intuitive and critical data

elements. Basically, the data description of each natural

disaster can not be separated from images, so image

recognition is an indispensable function of natural

disaster identification.

• Can cope with the complexity of natural disaster images：

The image elements related to natural disasters are

complex and professional. Sea land boundary,

administrative region division, etc. are the basic

elements to describe the geographic information of

natural disasters. Different colors represent the severity

of natural disasters. Complex geographic information

will have the superposition of longitude and latitude

information. The convolution algorithm has been

developed and applied in image recognition and

processing, so the application of convolution algorithm

can be more conducive to the recognition and analysis

of weather data of natural disaster images.

A. The Key Points and Processes of Crawler Technology

Related to Disaster Weather News Data

Taking the strong convective weather as an example, shown

as figure 2, the collection process of strong convective weather

news data based on Internet crawler is as below:

1) Retrieve the strong convection data in the target area on

the official public data platform, generate the initial

URL, and taking the strong convection forecast from

China Weather Network

http://products.weather.com.cn as example, regard it as

initial URL. meanwhile associate and store the relevant

URL in the database.

2) Keep to retrieve based on new URL, key words of

relevant areas, date and other relevant data and website

of strong convection.

3) According to the completion of database combed and

issued by news, set the conditions of judgement. If meet

the condition, then terminate the crawler.

4) Classify images, data and others based on the algorithm

and the layout logic of news required for strong

convection weather forecast

1) Association Degree and Reliability of Association Rules

In the mining of meteorological disasters, early warning and

other related website data published by portals and official

websites on the Internet, an essential content factor is the

connection between various of data. And it is required to

segment and summarize these algorithms. Assume that the

meteorological disaster related data retrieval set we apply is

I={i1,i2,⋯,im}, D is the database during disaster, then we can

indicate any meteorological disaster event by applying (ID, T).

ID is the corresponding event number, T={i1,i2,⋯,in} represents

the data related to early warning of natural disaster. Therefore,

we have three definitions as below which is according to the

content of natural disasters:

Definition 1. The association degree of association rules,

representing that the ratio relationship between the natural

disaster data set containing both X and Y and the total natural

disaster data set. The expression is shown as formula (1):

sup(→) = |{: ∪⊆, ∈}|/|D| (1)

Definition 2. The reliability of association rules,

representing the ratio relationship between the natural disaster

data set containing both X and Y and the natural disaster data

set only containing X. The general expression is as is shown as

formula (2):



= (→) = |{: ∪⊆, ∈}|/|T: X ⊆, ∈|

)

efinition 3. When the association degree of project

collection support (X) is more than minimum value, then this

weather disaster data set is regarded as the data set of frequency

natural disaster.

2) Bayesian Network Algorithm Model Based on Natural

Disaster Data Analysis

For the Bayesian network algorithm model, we apply it to

the relevant models of natural disaster data digging. It is not hard

for us to promote it to natural disaster data digging according to

chain rule of probability(shown as formula (3)).

P(X

,…

∏

(



|

, 

, … 

−1

)



=1

)

Among of it, x is the smallest non independent subset related

to natural disasters, and parent (Xi) ⊆ {x1, X2,..., Xi - 1} is as

follows: P(Xi|X1,X2,⋯,Xi -1)=P(Xi|Parent(Xi)). When we are

230

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 08:10:07 UTC from IEEE Xplore. Restrictions apply.

alking about variable assignment, it can be promoted to

Bayesian network joint probability distribution based on natural

disaster data mining [7],which is shown as formula (4).

,…

∏

(



|

(





(

−1

)))



=1

∏







|

(

(



))



=1

)

e can design the Bayesian network retrieval process of the

Internet such as natural disaster portal and official website data

release:

1) Confirm the value scope of relevant scientific data of

natural disaster, such as the radar reflectivity, rainfall,

etc. when strong convective weather occurs, and the

wind speed value of typhoon weather and others. Then

classify according to specific characteristics.

Confirm the reliability relationship between different

data variables, which is used to build structural

schematic with direction but without ring, so as to build

the Bayesian Internet structure, that is: based on a

certain kind of logic order, make variable X satisfy the

condition of Parent(Xi) (i=1 ,2 ,… ,n).

2) Regulate or obtain local probability through learning

P(xi|Parent(Xi)) to confirm relevant parameters of

Bayesian model of natural disaster [7].

B. Image Recognition and Classification Algorithm of

Disaster Weather Warning Based on Convolution Network

1) The Algorithm Based on AlexNet Model’s Convolution

Network

There is 11 layers of convolution network in total in the deep

learning model brought up by Alex Krizhevshy, including

convolution layer, pooling layer and full connection layer.

The main steps of convolution and pooling are as below:

=

W−F+2

+1 (5

)

=

W−F

+1 (6

)

pplied to the deep learning related to disaster weather

images, we can obtain the AlexNet size of feature matrix based

on disaster weather convolution and pooling: W means the

height and width of feature matrix of disaster weather data, F

means the size of the convolution core /pooling core of disaster

weather images, P represents the number of disaster weather

feature matrix plus 0, and S means the step size of convolution

or pooling of disaster weather images. If S<=1, then formula (5)

works. If S>1, then formula (6) works [8].

2) Algorithm Based on ResNet Network

ResNet network is a neural network brought up in 2015.

Compared with traditional neural network structure (series

stacking of convolution layer and full connection layer

according to the simple logic), and along with the deepening of

the network, and in order to avoid performance degradation, the

concept of residual learning is added, meaning “cross layer

connection” is stacked. Shown as below figure 3, the input of

Internet structure is X, output of F(X)+X, which avoids the

degradation of neural network to a certain extent.

Below are the two typical ResNet structures as BasicBlock

structure and BasicNect structure:

Shown as figure 4, the BasicBlockstructure is the scenario

with a shallow network, composed by two 3*3 convolution

layers. The number of paths is 64.

Shown as figure 5, the BasicNect structure is suitable for the

scenario with deeper network. First is 1*1 convolution layer

with the number of paths as 64. And next is 3*3 convolution

layer with the number of paths as 64. The last is 1*1 convolution

layer with the number of paths as 256 [8].

C. Image Recognition and Classification Application of

Disaster Weather Warning

Image recognition and classification of disaster weather

warning based on convolution network and shown as figure 6,

we take the strong convective weather, geological disaster, dry

weather and typhoon tracks as typical disaster weather cases.

• Train the images of disaster weather, analyze and

extract the feature values of various of disaster weather.

Most of the disaster images are classified by color

representation, therefore, the extraction of color feature

is one of the most important standards.

For example, the amount of precipitation of strong

convection, the green, blue and pink color are taken as major

colors to distinguish the amount of precipitation. In the training

process of precipitation related samples, firstly, it is needed to

read the classification and benchmark of the basic image palette

(different publishing platforms, different countries and regions

will have different color representations of strong convective

weather).

The typhoon track in the national precipitation forecast map

describes the relevant positions in yellow and the longitude and

latitude of the map, and obtains real-time information of the

typhoon track by analyzing the corresponding colors.

Dry weather is described by white, yellow, orange, red and

brown as no drought, light drought, medium drought, heavy

drought and special drought.

Geological disasters are described in red, orange and yellow

as extremely high risk, high risk and higher risk.

tart

eight Layer

F(X)+X

Relu

Identity

ig. 3. ResNet structure

231

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 08:10:07 UTC from IEEE Xplore. Restrictions apply.

tart

3*3，6

Relu

64-d

ig. 4. BasicBlock structure

tart

1*1，6

3*3，6

Relu

256-d

1*1，2

Relu

ig. 5. BotteNeck structure

• C

arry out the training initialization on disaster images

based on the extracted disaster weather feature data, the

details include:

o initialize the structure and parameters of

strong convection, set the weight and

threshold value

o initialize the structure and parameters of

geographical disaster, set the weight and

threshold value

o initialize the structure and parameters of dry

weather, set the weight and threshold value

o initialize the structure and parameters of

typhoon track, set the weight and threshold

value

• Carry out training based on existing feature and target

values, including

o Adjust the relevant threshold of strong

convection and calculate the error value

o Adjust the relevant threshold of geographical

disaster and calculate the error value

o Adjust the relevant threshold of dry weather

and calculate the error value

o Adjust the relevant threshold of typhoon track

and calculate the error value

• Firstly, sort out the results after training the network,

and then test network respectively by strong convection,

geological disasters, dry weather and test images of

typhoons.

• According to the test results, output the corresponding

strong convection identification results, geological

disaster identification results, dry weather identification

results and typhoon path identification results.

Above is typical disaster weather related research, which can

be extended to other related disaster warning, such as blizzard

forecast, hail forecast, haze warning and so on.

1) Application of Convolutional Neural Network Based

Disaster Weather Warning Classification and Recognition

Selection basis of sample data during the empirical analysis:

Firstly, the data of disaster weather are selected as samples

for analysis for two reasons: firstly, the data and images related

to disaster weather are more common, typical and representative.

Secondly, the types of weather are more complicated, such as

strong convection, morning fog, drought, etc. The samples are

richer and can be used as typical samples for crawlers and neural

network training. In refining the samples, that is, in selecting the

samples of disaster weather types, we cited the most common

disaster weather samples, that is, strong convective weather

(precipitation), dry weather, geological disasters, and typhoons.

Regarding the choice of training model structure:

convolutional + fully connected network) or using the classical

ResNet network, this training to typical convolutional neural

network as the main method, the specific model structure is

shown as Figure 7.

After three layers of convolution, the feature maps are

flattened and fully connected to output the score values of the

respective classification results, with the largest score being the

predicted category.

Since the main sample data is image data, our main model

structure is the typical convolutional neural network (based on

Training outcome:

We apply convolutional neural network to classify and

identify disaster weather warnings by using four types of

disaster weather warnings: strong convective weather

(precipitation), dry weather, national geological disaster, and

typhoon as experimental samples, the partical training daily

record is shown as Figure 8.

232

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 08:10:07 UTC from IEEE Xplore. Restrictions apply.

Training

network

djust the relevant threshold of

strong convection and calculate

the error value

tart

eature extraction of image

parameters of strong convective

weather

eature extraction of image

parameters of typhoon paths

eature extraction of image

parameters of dry weather

djust the relevant threshold of

dry weather and calculate the error

value

Data analysis

and image

extraction of

disaster

weather

features

djust the relevant threshold of

geological disaster and calculate

the error value

djust the relevant threshold of

typhoon paths and calculate the

error value

est the network after training

data output

nitialize the structure and

parameter of strong convection ,

set weights and thresholds

nitialize the structure and

parameter of geological disaster,

set weights and thresholds

nitialize the structure and

parameter of typhoon paths, set

weights and thresholds

nitialize the structure and

parameter of dry weather, set

weights and thresholds

Training

initialization

of disaster

images

ain the disaster weather

images

eature extraction of image

parameters of geological disaste

esult output of strong

convective image recognition

esult output of geological

disaster image recognition

esult output of disaster weather

image recognition

esult output of typhoon paths

image recognition

…

ig. 6.

The flow chart of images recognition and classification of disaster weather warning based on convolution network

233

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 08:10:07 UTC from IEEE Xplore. Restrictions apply.

ig. 7. ResNet structure

ig. 8. Partical training daily record

D. A

nalysis and Conclusion

Through the algorithm study in 3.1-3.3, as well as the

training and analysis with disaster weather as the sample, we can

conclude that the data of disaster news reports are related to

images.

(1) The crawler algorithm logic for early warning and real-

time data of disaster images can be further analyzed according

to the specific time date and geographical area in addition to the

type name of the disaster, so that real-time and predicted data

and images can be acquired for disasters in any administrative

area, and the timeliness of the material is improved, which helps

the rapid and automatic generation of news reports.

(2) By training a typical convolutional neural network for

disaster weather, we train and classify different disaster weather,

and by the relevant disaster weather features (e.g., strong

convective weather names as well as color intensity), we can

quickly classify the disaster weather. Extending to the

classification and recognition of other types of disaster-related

data and images, we can perform relevant training to achieve

artificial intelligence recognition of different disasters based on

standardized title names as well as image features.

(3) Absolutely, the crawler algorithm and typical

Convolutional neural network training have some limitations

related to disaster news reporting. First of all, the crawler

algorithm, we can only have access to publicly available

weather-related website data, for overseas data, unpublished

related weather information, we cannot get. On the other hand,

our typical convolutional neural network based recognition

classification also has some limitations, for example, the color

standard (palette) of disaster images in different countries will

be different to some extent, so that the classification recognition

will encounter related problems when dealing with publicly

available disaster warning and real-time images from different

countries, so we can only recognize training and classification

for images with consistent disaster image palette.

IV. F

URE

ELOPMENT

HAL

LENGES AND

PORTUNITIES

. Challenges: Artificial Intelligence Abuse

1) Law Risk

Along with the social development and progress of

technology, the artificial intelligence will gradually replace

many human positions, and the human reply more and more on

artificial intelligence. As a kind of high-tech, the damages

brought by illegal and criminal activities carried out by using

artificial intelligence technology are great. If these behaviors

and activities are not stopped by law in time, it will become one

of the factors leading social instability. The artificial intelligence

technology has been gradually rooted in many fields, resulting

in more and more ethical and legal disputes. For example, in the

production of data news report, the arguments on the ownership

of copyright and right of author of artificial intelligence are

never stopped.

In recent years, the shortcomings of artificial intelligence

technology have been gradually emerged in the news practice,

and the pursuit of news production subject to interests is

increasing day by day. The shortcomings of artificial intelligence

and supervision regulation of press field, and the incomplete

ethical system of news profession have all led to the prominence

of new ethics and legal issues in the process of combining

artificial intelligence with journalism. As a consequence, we

need to re consider the solution of news ethics [9].

2) Deep Forgery and Ethical Issues

234

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 08:10:07 UTC from IEEE Xplore. Restrictions apply.

eep forgery is a forgery technology generated on the

foundation of artificial intelligence deep learning, which mainly

forges other people's facial expressions and voices in real time

by using of artificial intelligence technology, and combines it

into new videos. In December,2017, a user named “deepfake”

posted a pornographic video impersonating a well-known

Hollywood actress on a foreign website “Reddit”, instantly

activating a carnival on the Internet. Deep forgery has been

applied in the film and television industry a few years ago,

creating special efforts scenes and makeup to achieve more

refined communication effects. However, at the same time, it

also provides opportunities for people with ulterior motives [10].

B. Opportunity: Overcome Difficulties

1) Automatic Writing of Disaster Weather News Based on

Artificial Intelligence

Based on current deep learning technology, we can sort out

and recognize the real-time disaster warning data and carry out

a series of classification. However, if edit these materials into

formal news that can be output directly so as to realize automatic

writing, a certain of technical difficulties still exists. Below is

mainly to carry out analysis in two perspectives:

• News writing has a certain degree of subjective thinking

and opinions of the writer, so as the disaster weather

news. Except information related to disaster warning, it

may still contain the prevention measurements, notices,

and subjective opinions under detailed situation. Taking

the most common strong convective weather as the

sample, except reporting the forecast, strength and

others of rainfall, it still needs to prepare some special

prevention measurements combined with the terrain,

topography, crowd density, morning and evening peak

travel of a specific area. All of these have opinions and

evaluation of subjective thinking. Therefore, the writing

of disaster weather news still needs a certain degree of

artificial assessment and modification.

• Achieve the technical challenges needed by automatic

writing. Based on above analysis, we know the key

point of realizing automatic writing is to realize the

consideration of factors with subjective thinking

patterns, which involve in humanities, environment,

geography, politics and other elements. Therefore,

sufficient data samples are needed, and then an

algorithm that can simulate news schemes with

appropriate subjective thinking should be available. So

the news writers can be completely replaced and editing

of news.

2) Timeliness of Disaster News

As we know, the special feature of disaster early warning

news and other news is that they have a certain of timeliness.

Due to it is involved in paying attentions to the safety of relevant

groups, and leaving enough time to make corresponding

prevention measures, therefore, it certainly has a request of

timeliness of disaster news.

First of all, with the help of artificial intelligence, compared

with traditional news writing, the semi self-help mode of early

warning of disaster weather news has been certainly increased

efficiency. However, with the special timeliness request of

disaster weather, it still faces the challenge of improving

efficiency. The details include below aspects:

• Improve the efficiency of computer’s calculation

efficiency and performance. Due to a huge amount of

historical data and samples is needed, deep learning of

disaster news requires large amounts of calculation

resource. Therefore, in the perspective of test and

business operation, both need higher amount of

calculation resources.

• Optimize storage structure and logic. For example,

based on the advantages of cloud calculation, apply

objective storage to separately store static data and

dynamic data. Meanwhile, realize the separation of read

and writing.

• Due to the disaster weather forecast has a request of

timeliness even to minutes and seconds, how to

optimize the algorithm and logic of deep learning, how

to quickly obtain sample data and complete initial

writing of news in the soonest time are the challenges

that need to be consistently optimized and improved.

ERENCES

] H. Chen, and L. Zhou, “From Wenchuan earthquake to Jiuzhaigou

earthquake: An analysis of changes in disaster news coverage,”

Journalism, no. 11, pp. 35-38+57, 2017, doi:10.15897/j.cnki.cn51-

1046/g2.2017.11.006.

[2] Q. Guan, M. Diao, and G. Q. Yao, “Teaching and practice of “intelligent

computing” course integrating geoscience content,” Chin. Geol. Educ., no.

3, pp. 38-42, 2021, doi:10.16244/j.cnki.1006-9372.20211014.002.

[3] J. Zhang, K. Xue, Yang, Zhipeng, F. Zhang, R. Zhang, J. Yang, and G.

Feng, “Application of artificial intelligence and internet of things in

atmospheric science,” Adv. Geophys., no. 1, pp. 94-109, 2022.

[4] L. Zhou, “Disaster governance for the era of artificial intelligence - a

multi-case based study,” Chinese Adm., no. 8, pp. 66-74, 2019,

doi:10.19735/j.issn.1006-0863.2019.08.08.

[5] T. T. Zhao, “Agenda setting of the media in natural disaster reporting,”

Youth J., no. 3, pp. 38-39, 2015.

[6] [6] H. Y. Wang, “On the innovation of financial media report of the

party newspaper in the event of natural disasters -- Take the new media

report of flood fighting and rescue in 2020 of Anhui Daily as an example,”

News World, no. 5, pp. 33-36, 2021, doi:10.19497/j.cnki.1005-

5932.2021.05.12.

[7] D. Shan, and Y. Shi, “Research on the application of data mining

technology in Internet information retrieval,” Sci. Technol. Bull., no. 3,

pp. 161-164, 2014, doi:10.13774/j.cnki.kjtb.2014.03.037.

[8] J. Jin, and P. Wang, “Research on image recognition algorithm based on

convolutional neural network,” Commun. Inf. Technol., no. 2, pp. 76-81,

2022.

[9] W. Niu, “The challenges of artificial intelligence technology on

journalism ethics and response strategies,” Master's thesis, Hebei

University of Economics and Business, Shijiazhuang, 2019.

[10] H. Li, “The criminal sanction of misuse of personal biometric information:

The example of artificial intelligence “deep forgery”,” Polit. Law Forum.,

no. 4, pp. 144-154, 2020.

235

Authorized licensed use limited to: Hong Kong University of Science and Technology. Downloaded on July 19,2023 at 08:10:07 UTC from IEEE Xplore. Restrictions apply.